LSB steganalysis using support vector regression

We describe a method of detecting the existence of messages, which are randomly scattered in the least significant bits (LSB) of both 24-bit RGB color and 8-bit grayscale images. The method is based on gathering and inspecting a set of image relevant features from the pixel groups of the stego-image, whose similarities and correlations change with different ratios of LSB embedding. The proposed detection scheme is based on support vector regression (SVR). It is shown that the measurement of a selected set of features forms a multidimensional feature space which allows estimation of the length of hidden messages embedded in the LSB of cover-images with high precision.